Back to Projects

Sembara

A cultural gamification ecosystem powered by a Serverless RAG Chatbot using Cloudflare Workers and Gemini API.

About this Project

Sembara Mobile Application Interface
The Sembara Mobile Application Interface.
Indonesia is home to thousands of cultural artifacts, yet learning about them often feels disconnected from the modern digital experience. Furthermore, local cultural practitioners (artisans, performers) struggle to find economic footing in the digital age.

Sembara bridges this gap. We created a mobile gamification ecosystem that makes cultural literacy addictive. Users explore a virtual map, complete cultural quests, and interact with a smart AI guide. Crucially, the app connects users directly to local artisans, turning cultural appreciation into economic empowerment.

This project was selected as a Top 50 National Finalist in the BudayaGo 2025 competition (Student Category).

System Architecture

The overall system architecture of Sembara's gamified cultural ecosystem.

The Sembara ecosystem consists of three main components:

  • Application
    Built with Flutter for high-performance rendering and a seamless gamified experience on Android.
  • Database
    Powered by Supabase for real-time synchronization of player stats (EXP, Achievements) and secure authentication.
  • AI Chatbot
    Built on Cloudflare Workers to run the RAG pipeline at the edge, minimizing latency.

My Role

As the AI Engineer, my goal was to create a chatbot that knows the cultural context without hallucinating, all while running with minimal latency.

The AI logic runs on Cloudflare Workers, The Worker handles payload construction, context injection, and API routing, effectively keeping heavy logic off the client device.

User Dashboard
The AI Chatbot interface within the Sembara mobile app.

To make the conversation feel natural, I implemented Cloudflare KV (Key-Value Store) as a high-speed, non-relational memory bank. When a user chats, the Worker pulls the chat history from KV, appends the new prompt, sends it to the AI, and then truncates & saves the updated history back to KV. This ensures that specific character personas are maintained seamlessly across sessions.

User Dashboard
The indexing and querying process of File Search in Gemini's RAG architecture.

We use Google’s Gemini API, tailored via Prompt Engineering, to maintain consistent cultural personas. To ensure accuracy, I used Gemini’s new Retrieval-Augmented Generation (RAG) feature. This involves offline indexing, where raw cultural documents are transformed into vector embeddings and stored in a vector database. During real-time querying, the system scans the database for relevant “chunks” of text which are fed into Gemini as context, allowing the AI to answer specific cultural questions accurately based on verified data.